1. Parallelizing Video Transcoding Using Map-Reduce-Based Cloud Computing Speaker : 童耀民 MA1G0222 Feng Lao, Xinggong Zhang and Zongming Guo Institute of Computer Science & Technology Peking University, Beijing 100871, P.R. China Email: {laofeng, zhangxg, guozongming}@pku.edu.cn 2013.07.24

2. Outline 1.INTRODUCTION 2.SYSTEM ARCHITECTURE 3.PROBLEM FORMULATION 4.MAX-MCT ALGORITHM 5.EXPERIMENT 6.CONCLUSION 2

3. INTRODUCTION  Recent years, there has been a growing demand for high quality video, which leads to advances of coding technology, such as H.264, MPEG-4 and MPEG-2 and so on.  And various environments usually require different coding formats. 3

4. INTRODUCTION  This results in the demand of fast transcoding.  However, due to the complexity of video coding, fast transcoding remains a problem to be explored. 4

5. INTRODUCTION  There have been many efforts devoted to parallel transcoding over multi-core processor, such as [1] [2] [3].  But due to specified hardware, the parallel transcoding over multi-core processor is hard to extend. 5

6. INTRODUCTION  Cloud computing, as an emerging technology, can utilize computing power of thousands of computers.  Cloud computing consists of a cluster of distributed computers. 6

7. INTRODUCTION  Since the computers can be heterogeneous, cloud computing is extendable and relatively inexpensive.  Map/Reduce is a distributing cloud computing model. 7

8. INTRODUCTION 8

9.  Moreover, when the transcoding time is in proportion to segment complexity, Min- Min algorithm is equal to minimal complete time (MCT) algorithm.  The MCT algorithm assigns segments according to descending complexity order. 9

10. INTRODUCTION  We formulate the scheduling as an NP-hard problem.  Considering overhead to launch sub-tasks, we propose a heuristic task scheduling algorithm, named Maximizing Minimal Complete Time (Max- MCT), which includes two procedures: virtual knapsack and MCT procedures. 10

11. SYSTEM ARCHITECTURE 11

12. SYSTEM ARCHITECTURE  To insure the independency of the segments, video sequence should be divided in between GOPs.  Moreover, the content of each segment is also various.  Therefore, the complexity of segments is heterogeneous. 12

13. PROBLEM FORMULATION  As present above, we are given n different segments, J = (1,2,3…..n)  with different complexity, C = (C 1,C 2,C 3 ….Cn)  Each segment must be processed without preemption until its completion.  We also have m computers with different capacity, P = (P 1,P 2, P 3 …P N )  And the task-launching overhead is T overhead. 13

14. PROBLEM FORMULATION 14

15. MAX-MCT ALGORITHM 15

16. MAX-MCT ALGORITHM 16

17. MAX-MCT ALGORITHM  B. MCT procedure  Here, we employ MCT algorithm to handle the residual pieces.  Then the computer with the minimal complete time is chosen.  It continues until all the residual segments have been assigned. 17

18.  C. Algorithm Analysis  Now, we analyze the complexity of our Max-MCT algorithm.  Sorting the segments according has complexity O(n log n).  The virtual knapsack procedure has complexity O(n) and the complexity of MCT algorithm is O(nm).  So our algorithm has a low complexity O(n log n). 18

19. EXPERIMENT  Thus we employ Matlab to conduct simulation experiments to evaluate different scheduling strategies.  Generally, we create 8 computers, with capacity ranging from 5 to 15, and 300 video segments whose complexity ranges from 300 to 900.  And the task-launching overhead is 10. 19

20. EXPERIMENT  For each situation, we conduct 1000 experiments and pick the average as an output.  Here we mainly evaluate the Max- MCT algorithm against MCT algorithm. 20

21. 21

22. 22

23. 23

24. 24

25. CONCLUSION  In this paper, we investigate the fast transcoding problem and present a Map-Reduce-based cloud transcoding system.  To reduce complexity, we propose a heuristic algorithm named Max-MCT with two procedures.  We also conduct various simulation experiments to verify that our algorithm outperforms the exiting algorithms. 25